Belin Doc IconBelin Doc
Released Jul 9, 2026 · Three capability tiers

GPT-5.6 Document Translation

GPT-5.6 is one generation sold as three capability tiers — Sol, Terra and Luna — priced five times apart. We translated 8 document scenarios with every tier, three runs each, and had two blind judges score the output. Sol averaged 4.70 and reached a perfect 5.00 in five of the eight scenarios; Terra was the fastest model in the whole field; Luna was the cheapest per point of quality. Belin Doc offers Sol and Terra today.

Highest quality
4.70GPT-5.6 Sol
Fastest
3,598 msGPT-5.6 Terra
Cheapest per quality point
$0.00026GPT-5.6 Luna

Sol, Terra and Luna — measured, not marketed

The previous-generation GPT-5.5 is included as a baseline, because Sol carries an identical list price and the comparison is the one most people actually need.

Scores are the median of 3 runs across 8 document scenarios, blind-judged by GPT-5.4 and Claude Opus 4.7 (neither judge is a contestant). Latency was measured sequentially at concurrency 1. Prices are the official standard tier, verified 2026-07-10.
MetricGPT-5.6 SolGPT-5.6 TerraGPT-5.6 LunaGPT-5.5Previous generation
Overall score4.704.384.344.54
Scenarios scored a perfect 5.005 / 82 / 81 / 80 / 8
API price (input / output per 1M tokens)$5 / $30$2.50 / $15$1 / $6$5 / $30
Cost per call$0.00360$0.00161$0.00113$0.00393
Cost per quality point$0.00077$0.00037$0.00026$0.00087
Median latency4,640 ms3,598 ms7,998 ms3,925 ms
Best forHard long documents, legal and academic text, terminology consistencyLatency-sensitive and batch translationThe lowest cost per point of qualityPrevious-generation flagship, kept as a baseline
On Belin DocAvailableAvailableNot offeredAvailable

Which tier should you actually pay for?

  1. You need the best translation you can get

    Sol

    It reached a perfect score in five of the eight scenarios. It also costs the same list price as GPT-5.5 while scoring 0.16 higher and spending less per call, because it writes fewer output tokens to say the same thing.

  2. Latency is what you are optimising

    Terra

    At a median 3,598 ms it was the fastest of all seven models we tested, and it still scores 4.38 — comfortably above the open-weight alternatives.

  3. Cost is what you are optimising

    Terra, for now

    Luna has the lowest cost per quality point of the family, but it is not offered on Belin Doc yet. Among the tiers you can select today, Terra is the cheaper one.

One result that contradicts the marketing

Luna is positioned as the light, fast tier. It is indeed the cheapest — but in our sequential measurements it was the slowest of the three, at 7,998 ms, roughly 2.2 times Terra. Caveat worth stating: every model was called over the same request path, so part of that spread may reflect the path rather than the model itself.

How GPT-5.6 places against everything else we tested

The top three finish within 0.03 points of each other — inside judge noise. What separates them is price: Sol reaches that level at one of the two lowest costs per quality point in the field, while Gemini 3.1 Pro costs about 3.4 times more per point.

Seven models, eight document scenarios, three runs per cell, median taken. Overall score is the mean of five dimensions: faithfulness, fluency, terminology, style and formatting.
RankModelOverallCost / callCost / pointMedian latency
1Claude Opus 4.84.73$0.00366$0.000774,084 ms
2Gemini 3.1 Pro4.71$0.01222$0.002594,380 ms
3GPT-5.6 Sol4.70$0.00360$0.000774,640 ms
4GPT-5.54.54$0.00393$0.000873,925 ms
5GPT-5.6 Terra4.38$0.00161$0.000373,598 ms
6GPT-5.6 Luna4.34$0.00113$0.000267,998 ms
7DeepSeek V4 Pro4.18$0.00068$0.000167,372 ms

How it was measured

Models tested
7
Document scenarios
8
Runs per cell (median taken)
3
Translations judged
168
Blind judges, neither a contestant
GPT-5.4 · Claude Opus 4.7
Verified on
2026-07-10

Where Sol earns its score

Judges scored every translation on five dimensions, 1 to 5. These are Sol's medians across all eight scenarios.

Faithfulness
4.8

Nothing added, nothing quietly dropped — the hard requirement for contracts and research papers.

Fluency
4.8

Reads as if it were written in the target language, not decoded from the source.

Terminology
4.8

A defined term on page 3 is still the same term on page 80, across the whole document.

Style
4.7

Register survives the crossing: legalese stays legal, literary prose stays literary.

Formatting
4.5

Tables, code identifiers and Markdown structure come through intact.

Translate with GPT-5.6 in three steps

  1. 01

    Upload your document

    Drop in a PDF, Word, PowerPoint, Excel or EPUB file — whole books included.

  2. 02

    Pick Sol or Terra

    Choose the GPT-5.6 tier that matches your priority, then set the target language.

  3. 03

    Download the translation

    Layout, tables and terminology come back preserved, in the original file format.

GPT-5.6 document translation FAQ

What people ask before choosing a tier.

What is the difference between GPT-5.6 Sol, Terra and Luna?
They are three capability tiers of the same generation, not three generations. Sol has the highest reasoning ceiling and the highest price ($5 / $30 per 1M input / output tokens). Terra is the balanced tier ($2.50 / $15). Luna is the lightweight, cheapest tier ($1 / $6). In our benchmark Sol averaged 4.70, Terra 4.38 and Luna 4.34.
Which GPT-5.6 tier is best for document translation?
Sol, for anything where quality matters — it scored a perfect 5.00 in five of our eight document scenarios. Choose Terra when latency matters: at a median 3,598 ms it was the fastest of the seven models we tested, and it still scores 4.38.
Can I use GPT-5.6 Luna on Belin Doc?
Not currently. Belin Doc offers GPT-5.6 Sol and GPT-5.6 Terra. Luna was included in our benchmark for completeness, and we publish its results so the family comparison is honest, but it is not selectable as a translation model today.
Is GPT-5.6 better than GPT-5.5 for translation?
Sol is: it scores 4.70 against 4.54 at the same list price, and it costs less per call because it produces fewer output tokens. Terra and Luna score below GPT-5.5 on overall quality, but cost roughly a half and a third as much per call respectively — which is the trade they exist to offer.
How were these scores measured?
Seven models translated eight document scenarios — academic, legal, technical, literary, manga dialogue, Markdown tables, long-document terminology and more — three times each, for 168 translations. Candidate translations were anonymised and reshuffled each round, then scored 1 to 5 by two judges, GPT-5.4 and Claude Opus 4.7. Neither judge was a contestant, so no model graded itself. Every raw translation and judge score is archived in full, and the numbers reported here come straight from that archive.
How much does GPT-5.6 translation cost?
Measured on real documents rather than list price: Sol averaged $0.00360 per call and Terra $0.00161. Divided by the quality score they achieve, that is $0.00077 and $0.00037 per quality point. On Belin Doc you can start translating for free; larger documents and premium models may use credits depending on your plan.
Which file formats and languages are supported?
PDF, Word (DOCX), PowerPoint (PPTX), Excel, EPUB and more, with the original layout, tables and formatting preserved. Belin Doc supports translation across 100+ languages.
Is GPT-5.6 document translation free?
You can start translating for free. Free monthly pages are included on every account, and larger documents or premium models may draw on credits depending on your plan.

Sources

Every number on this page is traceable. The scores and latency come from measurements we ran ourselves; the prices are taken straight from the vendor’s official pricing page. Nothing here is quoted from a launch announcement.

SourceTypeVerifiedUsed for
Belin Doc in-house benchmark dataset7 models × 8 scenarios × 3 runsFirst-party measurement2026-07-10Scores, dimension medians, perfect-score counts, cost per quality point
Sequential latency captureconcurrency = 1First-party measurement2026-07-10Median latency, measured sequentially at concurrency 1
developers.openai.com/api/docs/pricingOpens in a new tabOfficial vendor pricing2026-07-10Official list prices for all three tiers
GPT-5.6 Sol vs Terra vs LunaPublished review2026-07-10Full method and the per-scenario breakdown of all eight scenarios

Only the last two rows link out: the pricing row goes to OpenAI’s official pricing page, and the review carries the full per-scenario breakdown. The first two rows name the raw measurement artifacts these numbers were computed from.